GLM 5.2 Review: Z.ai's Million-Token Workhorse

A 1,048,576-token context window, 131,072 tokens of output per response, MIT-licensed weights, and under a dollar per million input tokens. That spec sheet is why this release has been the one developers keep mentioning since Z.ai shipped it in mid-June 2026. This GLM 5.2 review covers what the model actually is, what the million-token window is good for, and whether it deserves a spot in your rotation — or just a bookmark.

Spoiler: if you work with large codebases or long documents, it deserves the spot.

What Is GLM 5.2 and Why Does It Matter?

Z.ai (the company formerly known as Zhipu AI) has been shipping GLM models at a relentless pace, and this June release is the most consequential since GLM 5 itself. It is built for long-horizon, agentic software engineering rather than casual chat: the kind of work where a model needs to hold an entire repository, a sprawling requirements doc, and a long tool-use history in its head at once.

The architecture is a mixture-of-experts transformer — reported at roughly 753 billion total parameters with around 40 billion active per token — which is how it keeps serving costs at standard-tier levels despite frontier-scale capacity. Z.ai also shipped a dual thinking-effort system: two reasoning levels that let you trade latency and cost against depth, in the same spirit as the thinking modes we covered in our guide to when reasoning pays off. Coverage from MarkTechPost noted an unusual detail about the launch: Z.ai published no benchmark table at release, an oddly confident move that left the community to run its own evaluations.

One launch quirk worth knowing: the model arrived first for subscribers of Z.ai's GLM Coding Plan on June 13, with the MIT-licensed weights and standalone API access following on June 16 — which is the availability date you will see on most model catalogs, including CoreAI's.

How Big Is a Million-Token Context, Really?

Numbers this large stop meaning anything without translation, so: 1,048,576 tokens is roughly 750,000 words. That is an entire mid-sized codebase — not a file, the codebase — or about ten novels, or a year of meeting transcripts, in a single prompt. GLM 5.1 topped out around 200K tokens; the new release quintuples that.

The jump matters most for agentic work. Long agent sessions die when the context fills up and the model starts forgetting what it did an hour ago. A five-times-larger window means five times longer before that wall, and Z.ai paired it with an architectural optimization (reported as IndexShare) that cuts per-token compute at extreme context lengths — because a million-token window you cannot afford to fill is a spec-sheet trick, not a feature.

The 131,072-token output cap is the quieter half of the story. Most models cap responses at 16K-64K tokens, which forces multi-file refactors into many turns. Z.ai's model can emit a large multi-file diff — or a genuinely book-length document — in one shot. If long context is the axis you buy models on, our comparison of 1M-token context models shows how its peers stack up.

What Does GLM 5.2 Cost Compared to Its Rivals?

Here is the model next to the rivals it actually competes with, using per-million-token rates from CoreAI's live catalog:

The positioning is clear: Z.ai charges less than half of Claude Sonnet 5's input rate and under a third of its output rate while matching the million-token window. Qwen3.7 Plus and MiniMax M3 are cheaper still, but neither matches this combination of window size and 131K single-shot output. Cached input drops to about $0.17 per million tokens, which matters enormously for agent loops that re-read the same repository state every turn.

Where does it sit in the GLM family? Think of it as the long-context flagship above the models we broke down in GLM 5 Turbo vs GLM 5 vs GLM 4.7 Flash — if you read that guide and wanted GLM 5's quality with a dramatically bigger window, this is the answer.

When Should You Actually Use GLM 5.2?

Reach for it when the problem is bigger than a file. Repository-scale refactors, code review across a whole service, migrating a legacy module while keeping the rest of the system in context, digesting a 400-page spec and producing an implementation plan — these are tasks where smaller-window models need elaborate retrieval scaffolding and this one just reads everything. The huge output cap also makes it one of the best models available for generating long structured documents: full API references, complete test suites, book-length reports.

It is also a statement in the broader story we covered in how Chinese AI models are winning enterprise share: an MIT-licensed, frontier-window model at standard-tier prices is exactly the pattern that has made DeepSeek, Qwen, and GLM impossible for cost-conscious teams to ignore. The open license means a team can prototype against the hosted model and later self-host the identical weights — no vendor lock-in cliff.

Where would I not use it? Quick everyday questions — a million-token window buys you nothing on "fix this regex" and a budget model will answer faster and cheaper. And for multimodal work it is the wrong tool: the model is text-in, text-out, so screenshots and diagrams need a different model. Since Z.ai published no launch benchmarks, the sensible move is to test it yourself: paste a real task into CoreAI Compare and run it head-to-head against Claude Sonnet 5 or Qwen3.7 Plus. Your own prompt is the only benchmark that counts.

How to Try GLM 5.2 Without an API Key

GLM 5.2 is live on CoreAI today across iOS, Android, and the web — one subscription covers it along with 300+ other models, no API keys or per-token billing to manage. Pick it from the model list, toggle thinking mode when you want the deeper reasoning effort, and attach files when you have long documents to feed that giant window. If you write with it regularly, CoreAI's free AI Outline Generator is a handy companion for structuring the long-form drafts this model is so good at filling in.

Frequently Asked Questions

Is GLM 5.2 open source?

Yes. The weights are released under an MIT license, one of the most permissive options available, continuing the GLM family's open-weight track record. You can run it yourself or use a hosted version; on CoreAI it is available as a standard-tier model with no setup.

What are the context window and output limits?

The context window is 1,048,576 tokens — about 750,000 words — with up to 131,072 tokens of output in a single response. That is roughly five times the context of GLM 5.1 and an unusually large output cap, big enough for multi-file code diffs or book-length documents in one generation.

How much does GLM 5.2 cost?

On CoreAI, the model runs about $0.93 per million input tokens and $2.93 per million output tokens, with cached input around $0.17 per million. That is less than half the input price of Claude Sonnet 5, the model whose context window it matches.

Is GLM 5.2 good for coding?

That is its home turf. Z.ai positions it for long-horizon agent workflows and project-level software engineering — repository-scale refactors, service-wide code review, and long agentic sessions that would overflow smaller context windows. For single-file quick fixes, a cheaper model is usually the smarter pick.

Does GLM 5.2 support images or other modalities?

No. GLM 5.2 is text-in, text-out. If your workflow needs screenshots, diagrams, or video input, pair it with a multimodal model — on CoreAI you can switch models mid-conversation, so mixing it with a vision-capable model in one thread is straightforward.